Electronic Theses and Dissertation

Penelitian ini mengembangkan sistem klasifikasi otomatis untuk membedakan larutan
H₂O murni dan H₂O campuran NaOH menggunakan citra multispektral pola High-High
Fluctuation (HHF). Secara visual kedua larutan tampak identik sehingga diperlukan
pendekatan berbasis pengolahan citra digital. Metode yang diterapkan meliputi tiga tahap: (1)
segmentasi K-Means untuk mengekstrak area larutan dari latar belakang, (2) klasifikasi
menggunakan tiga arsitektur Convolutional Neural Network (CNN) yaitu InceptionV3,
VGG19, dan DenseNet201, serta (3) penggabungan hasil klasifikasi melalui Auto Weighted
Decision Fusion. Dataset terdiri dari 1.050 citra (525 H₂O dan 525 H₂O+NaOH) yang dibagi
menjadi 70% data latih, 15% validasi, dan 15% pengujian. Hasil eksperimen menunjukkan
bahwa VGG19 mencapai akurasi pengujian tertinggi sebesar 98,55% dengan F1-score 0,98
pada kondisi evaluasi dengan segmentasi K-Means. Namun, penentuan bobot pada tahap Auto
Weighted Decision Fusion didasarkan pada evaluasi independen tanpa segmentasi K-Means,
yang mengukur ketangguhan (robustness) masing-masing arsitektur secara mandiri. Pada
kondisi tersebut, DenseNet201 menunjukkan akurasi tertinggi sebesar 95,86%, diikuti
InceptionV3 sebesar 82,82%, dan VGG19 sebesar 46,58%, yang mengindikasikan bahwa
VGG19 memiliki ketergantungan tinggi terhadap kualitas input hasil segmentasi. Berdasarkan
hasil evaluasi independen tersebut, bobot yang ditetapkan adalah DenseNet201 (0,43),
InceptionV3 (0,37), dan VGG19 (0,21). Setelah penerapan Auto Weighted Decision Fusion,
sistem mencapai akurasi akhir sebesar 91,5%, yang mencerminkan peningkatan stabilitas dan
konsistensi prediksi dibandingkan model tunggal dalam menghadapi variasi distribusi data.
Sistem diimplementasikan dalam platform berbasis website sebagai antarmuka klasifikasi yang
dapat diakses secara langsung.

This study developed an automatic classification system to distinguish pure H₂O and H₂O mixed with NaOH using multispectral images with High-High Fluctuation (HHF) patterns. Visually, both solutions appear identical, necessitating a digital image processing-based approach. The applied method involved three stages: (1) K-Means segmentation to extract the solution area from the background, (2) classification using three Convolutional Neural Network (CNN) architectures: InceptionV3, VGG19, and DenseNet201, and (3) combining the classification results using Auto-Weighted Decision Fusion. The dataset consisted of 1,050 images (525 H₂O and 525 H₂O+NaOH) divided into 70% training, 15% validation, and 15% testing. Experimental results showed that VGG19 achieved the highest test accuracy of 98.55% with an F1-score of 0.98 under the evaluation conditions with K-Means segmentation. However, the weighting in the Auto-Weighted Decision Fusion stage was based on an independent evaluation without K-Means segmentation, which measures the robustness of each architecture independently. Under these conditions, DenseNet201 demonstrated the highest accuracy of 95.86%, followed by InceptionV3 at 82.82%, and VGG19 at 46.58%, indicating that VGG19 is highly dependent on the quality of the segmentation input. Based on these independent evaluation results, the weights assigned were DenseNet201 (0.43), InceptionV3 (0.37), and VGG19 (0.21). After applying Auto-Weighted Decision Fusion, the system achieved a final accuracy of 91.5%, reflecting improved stability and consistency of predictions compared to a single model in the face of varying data distributions. The system is implemented on a website-based platform as a classification interface that can be accessed directly.